8ight.AI, a Canadian company that sells services for appearing in AI answers, published a study on October 5, 2026 showing that ChatGPT and Claude read different websites when asked for a local tradesperson. According to 8ight.AI, Claude cited Yelp in 202 of its 218 answers, while ChatGPT cited it in none of 1,412.

In Short

8ight.AI asked two AI chatbots, ChatGPT and Claude, who the best roofer, plumber or similar tradesperson is in dozens of towns, then checked which websites each chatbot pointed to and which businesses it named. The two chatbots read very different places - Claude leaned on Yelp, ChatGPT on the Better Business Bureau and government licensing pages - and they often named different businesses, so a company can be recommended by one assistant and passed over by the other. The company that ran the test sells help with AI visibility, and it asked through developer connections rather than the apps people use, so the numbers describe a September 2026 snapshot instead of a fixed ranking system.

What the study measured

The study carries the title "Ask an AI for a roofer" and sits on the company's own website as the first entry in a research series. 8ight.AI put one question, word for word, to each assistant: "Who is the best {trade} in {city}? Name specific businesses." The questions went through each company's API with web search switched on, covering 39 trades in 49 places - 32 in the United States and 17 in the Greater Toronto Area - according to 8ight.AI. Answers were gathered between September 10 and September 30, 2026, although 1,391 of ChatGPT's 1,412 fresh answers and 199 of Claude's 218 arrived on September 10 and 11.

The yardstick was a list of 2,976 local businesses taken from Google Maps searches through DataForSEO's Google Maps data: 2,227 in the United States and 749 in the Greater Toronto Area, spread over 516 trade-and-place pairs, or about six businesses per pair. Only businesses with their own website and public contact details qualified. The study calls the group "a selection, not every business Maps shows", and it names none of them anywhere in the text.

A business counted as named when its name, or a plain short form of it, appeared in an answer written with capital letters and not running on into a longer name. A different branch of the same chain did not count. The rule exists as published code (matcher.py). A source, in the study's terms, is the website domain of a link that the engine returned with its answer.

Two engines carry the findings. ChatGPT supplied 1,412 answers, 1,398 of them from OpenAI's gpt-5.6-terra model and 14 from gpt-5.6-sol. Claude supplied 218: 200 from Anthropic's Claude Sonnet 4.6 and 18 from three other Claude models that 8ight.AI's checks used later in September. Gemini (gemini-2.5-flash) answered 20 times. That was too few for comparison, and because Gemini's links come back as Google redirect addresses, it is excluded from every source figure.

The thin Claude and Gemini counts trace to 8ight.AI's own credits with those companies running out, according to the study. Of 1,413 fresh questions, Claude answered 218 and Gemini 20.

One discrepancy sits between the documents. 8ight.AI's summary of the study describes an exercise with ChatGPT and Claude. The full study page lists Gemini as a third engine, then states that the findings rest on the first two.

Where each engine reads

The sharpest result is the split in sources. The table shows the share of fresh local answers that cited each platform at least once, according to 8ight.AI.

PlatformChatGPT (1,412 answers)Claude (218 answers)
Yelp0.0% (0)92.7% (202)
Better Business Bureau38.7%6.4%
Government or licensing site25.1%0.0%
Expertise.com16.3%22.0%
Angi11.8%25.2%
Thumbtack3.7%19.3%
HomeGuide0.1%23.4%
ClassPass (fitness only)1.3%30.3%
Yellow Pages0.2%12.4%
Facebook0.0%8.7%
Reddit0.4% (6)0.0% (0)
Google Maps or Google business pages0.2%0.5%

Grouped by kind of source, ChatGPT cited the website of at least one Google Maps business in 68.1% of its answers, against 45.0% for Claude. Claude cited a directory or review site in 99.5% of answers and ChatGPT in 65.3%. Social networks appeared in 0.6% of ChatGPT answers and 11.0% of Claude's.

Government sites are the unexpected entry. One ChatGPT answer in four cited one, and Claude cited none. 8ight.AI read every government address ChatGPT cited - 355 answers - and found that in about 7 of 10 (250 of 355) the page was a licensing board, a licence look-up or a list of registered or approved contractors. Most of the rest covered permit rules, consumer guides and regulations. None was a tourism page. The study treats its government figure as a floor, because the rule counts addresses ending in .gov plus a few named regulators, and a state licensing portal on another type of address goes uncounted. 8ight.AI notes that a lapsed licence or an old business name in a public registry would sit on the kind of page ChatGPT cites.

Google is mentioned far more often than it is linked. Its name appears in the text of 23% of ChatGPT answers and 43% of Claude's, yet neither engine linked to a Google page in more than 1% of answers. Reddit barely registered: it was cited in 6 of the 1,630 answers across both engines, all six from ChatGPT. On the recurring Toronto-area questions, ChatGPT cited it once in 1,354 answers and Claude never. A smaller test by the UK company Whito, run in July 2026 across three towns and 36 questions, found Reddit cited in 10 of ChatGPT's 18 answers, according to Whito, so the picture plausibly varies by market and method.

The study adds a boundary on what a citation means. It is a link shown with an answer, "not everything the model has read", and a model may have learned from Reddit or any other site in training without showing a link. All figures also describe API output with web search, which can differ from what the consumer apps display to a signed-in person with a location and a history, according to 8ight.AI.

The source layer for a different kind of question was the subject of another study examined by PPC Land on September 27. A Berlin agency that places comparison articles on media sites found that 25.4% of 106,758 English citations from ChatGPT, Google AI Overview and Perplexity, in answers about the best product in a category, led to pages with a disclosed commercial interest. The subject, engines and method differ from the 8ight.AI work, and each study comes from a company that sells services in the field.

Reviews, shortlists and the long tail

For ChatGPT, the count of Google reviews tracked the chance of being named. 8ight.AI grouped businesses by review count at the time they were found on Google Maps.

Google reviewsBusinessesNamed by ChatGPTShare
Under 25435153.4%
25 to 999039010.0%
100 to 24977510813.9%
250 to 99952711020.9%
1,000 or more1353022.2%

The median stood at 175 reviews for businesses ChatGPT named and 97 for the rest. Star ratings showed no matching pattern: the median was 4.9 for named and unnamed businesses alike. Review counts date from August or September 2026, when 8ight.AI found each business, not from the day of each answer, and they are on file for 2,775 of the 2,975 businesses ChatGPT answered for.

Volume did not guarantee a place either. In the 285 trade-and-place pairs where at least four businesses with a review count were checked, the most-reviewed one was named 22.8% of the time (65 of 285). The others were named 10.3% of the time (215 of 2,085). That is more than twice the rate, yet the leader was still "left out about three times in four", in the study's words. "Most reviewed" here means most reviewed among the businesses 8ight.AI checked in that place, not among every business on Google Maps.

Shortlists are short. ChatGPT named 369 of the 2,975 businesses it answered for (12.4%), Claude 198 of 1,776 (11.1%) and Gemini 17 of 232 (7.3%). 8ight.AI cautions against reading this as an AI that ignores nearly nine businesses in ten. The assistants return a short list - a median of at least five names from ChatGPT and seven from Claude, with storage capped at eight - while Google Maps returns a long one.

The chance that ChatGPT named at least one checked business in a place varied by trade. For trades with 12 or more places in the sample, Pilates studios led at 90.5% (206 businesses across 21 places), followed by women's fitness gyms at 80.0% and towing at 64.3%. Roofing contractors sat at 50.0%, plumbers at 47.4% (230 businesses across 19 places) and HVAC contractors at 45.0%. At the other end came house cleaning at 26.7%, barre studios at 16.7% and boutique fitness studios at 6.3%. The number of businesses checked per place ran from about 3 to 12 depending on the trade, which is why 8ight.AI asks for the gaps to be read as "a direction, not a ranking".

Two engines, two lists, and a different list on the second ask

Where ChatGPT and Claude answered the same question at the same moment, they rarely agreed. Across 1,775 businesses, 294 were named by at least one of them: 81 (28%) by both, 96 by ChatGPT only and 117 by Claude only. Of the 198 businesses that all three engines answered for, 29 were named by at least one engine and 2 by all three.

Comparing the lists of names themselves, including businesses outside the checked set, gave a similar result. Across 214 questions in 211 trade-and-place pairs, a median 12% of all businesses named by either engine were named by both, and for 24% of the questions the two lists had nothing in common. The stored lists stopped at eight names, and 164 of Claude's 214 lists and 110 of ChatGPT's reached that limit, so some answers named more businesses than were compared. A cleaning script dropped 18.9% of 11,393 stored entries as not being a business, such as an address, a directory or a summary line. A hand check of that script found two missed matches and some retained entries that were not businesses; correcting both patterns moved the median overlap from 12% to 11% and the no-overlap share from 24% to 23%. The published script is the uncorrected one.

Repetition produced further variation. In 267 trade-and-place pairs ChatGPT received the same question twice, a median 39 minutes apart, with 61% inside an hour. A median 29% of all businesses named in either answer appeared in both, and for 9% of pairs the two answers shared nothing. The sources behaved similarly. On consecutive runs a median one day apart, a median 29% of the sources ChatGPT cited in either answer appeared in both (1,073 pairs), while Claude, on a single model, reached 78% (805 pairs). Pairs a week apart on one model were few: 34 for Claude, with a median overlap of 67%, and none clean for ChatGPT.

Caveats attach to the repeatability figures. 8ight.AI did not store the model with each answer, so model periods were pieced together from its own deployment history, and it changed models, settings and tracked questions during the period. The stability figures therefore compare only consecutive answers from the same model period.

Instability in AI citations has surfaced in other measurements. SISTRIX's 17-week citation drift study found 89% of the non-core citation set in Google's AI Mode rotating weekly. That is a different product measured differently, though the direction is the same.

Limits, failures and hand checks

The study lists its own weaknesses at length. The data is a snapshot taken mostly on two days. An earlier answer for the same trade and place could be reused for up to seven days, and a reused answer was counted once. The sample covers two countries, 39 trades and mostly large US cities. It contains nothing on restaurants, retail, law or business-to-business buying, and in health care only med spas. Businesses without a website are absent, and 8ight.AI says they may be the least visible of all. The category "business website" is undercounted, since it recognises only sites of businesses found on Google Maps.

The failure log is part of the record. Across all scheduled AI checks run from July 1 to October 5, 2026, 37.5% of attempts failed (5,046 of 13,473), and 4,720 of those failures came from billing, quota or a spending cap on 8ight.AI's side. By engine, the failure rate was 67.7% for Gemini, 37.5% for Claude and 7.1% for ChatGPT.

On accuracy, 8ight.AI describes a replaced rule. Its first, looser test counted a business as named when every distinctive word of its name appeared in one sentence. A check of 50 random "named" verdicts found 13 wrong, mostly businesses whose names consist of common words. The stricter rule used for the study was checked on fresh samples: 39 of 40 "named" verdicts were right. Among 40 "not named" businesses sharing a distinctive word with the answer, it missed one or two that the AI had written in a different form, so the study expects its rates to be "a little low" rather than high, in its own phrasing. Hand checks of list matching covered 30 ChatGPT-and-Claude list pairs and 20 repeated-question pairs and found no false matches.

The recurring questions come from a separate set: 120 questions about local services in Toronto and nearby Ontario towns, 84 of them first asked on or after September 22. They yielded 2,527 ChatGPT and Claude answers between July 1 (ChatGPT from August 23) and October 5, 2026, and serve only for source stability and one Reddit count.

The company states its interest directly: it sells services that help businesses appear in AI answers, and so "we have an interest in this subject", in the study's wording. Exclusions are listed. Checks run by visitors to the company's free tool were left out, because they may contain personal data, as were one trade under a standing company rule (25 checks) and one trade-and-place pair to avoid a conflict of interest (24 checks). All queries were rerun for publication between 01:31 and 02:15 UTC on October 6, 2026, the evening of October 5 in Toronto. The study lists no corrections, and says "a correction is added here as a dated note" if one arises.

Why the findings matter to marketers

Two stages of AI answers now draw measurement attention: what crawlers take from sites, and what the finished answer cites. PPC Land's September 29 coverage of publisher lobbying noted that AI bot visits to publisher sites rose from one per 200 human visits in the first quarter of 2025 to one per 31 by the fourth, and that publishers can see crawlers such as GPTBot and ClaudeBot because those agents announce themselves. The 8ight.AI study sits at the output end. It records which domains appear as links beneath an answer, rather than which pages a bot fetched.

For ad and media buyers, the open question is how much of that output is organic and how much is bought. The study did not examine paid placements, and its answers came through APIs rather than the consumer apps. In the apps, PPC Land reported that Similarweb recorded sponsored placements in 26% of ChatGPT responses shown to Free and Go users in August. Whether those placements interact with the organic shortlists measured by 8ight.AI is outside what the data can say.

Measurement discipline is another thread. Only 16% of brands track AI visibility systematically, according to an IAB framework published on August 3, 2026, and the 8ight.AI repeat-question figures show why single checks vary: the same question put to ChatGPT 39 minutes apart, on median, produced lists sharing under a third of their names. The study itself frames a single check as one sample and lists the number of times a question was asked, the engine and the date as the details attached to any claim.

The commercial context deserves a plain statement. 8ight.AI works in the field often called GEO, and publishes figures in a form that can be checked, with each figure linked to its SQL or script. The company also stresses limits on interpretation. Its first reading of the data, stated in the study, is that it "shows where the engines read, not what moves them", and it says it did not test whether improving a profile on any listed platform changes an answer.

What the dataset establishes is narrow. In September 2026, two assistants asked identical questions produced different source sets, different shortlists and partly different lists on a repeat ask. Review volume lined up with being named for ChatGPT, star rating did not, and a business in a government licensing registry sat on a page type that ChatGPT cites often and Claude did not cite at all. Whether any of that moves a customer's choice is not measured.

Timeline

Summary

Who: 8ight.AI (8IGHT AI TECHNOLOGY INC.), a company based in Etobicoke, Ontario that sells services helping businesses appear in AI answers. The study covers ChatGPT (OpenAI), Claude (Anthropic) and, in minor volume, Gemini (Google), and compares them with 2,976 local businesses.

What: A study titled "Ask an AI for a roofer" that analysed 1,630 fresh answers from ChatGPT and Claude to the question of who the best tradesperson is in a given place. Claude cited Yelp in 202 of 218 answers and ChatGPT in none of 1,412; ChatGPT cited the Better Business Bureau in 38.7% of answers and government sites in 25.1%. ChatGPT named 15 of 435 businesses with fewer than 25 Google reviews and 30 of 135 with 1,000 or more, and the two engines named the same businesses in 28% of the cases where either named one.

When: Answers were collected between September 10 and September 30, 2026, recurring questions ran from July 1 to October 5, 2026, the study was published on October 5, 2026, and 8ight.AI sent a summary to PPC Land on October 6, 2026.

Where: 49 places in the United States (32) and the Greater Toronto Area (17), across 39 trades, asked through each company's API with web search enabled.

Why: The study measures which websites AI assistants link to when they recommend local businesses, a question that sits downstream of the crawler-side coverage PPC Land has run. For marketers and publishers, it gives a September 2026 baseline on source differences between engines, the weak link between star ratings and being named, and the instability of repeated answers, with the caveat that its author sells services in this field and that the data came through APIs rather than consumer apps.